x86_64: fix packedStore miscomp by spilling EFLAGS

[xtexx]

Fixes #20113. Fixes #20581. Fixes #22790.

AND instructions in packedStore clobbers EFLAGS.

Tested with the repro from #20113 and #20581.

xtex@xtex1 ~/s/z/zig (master)> build-host/out/zig run repro.zig -fno-llvm
ret: repro.Data{ .rd = 1, .rl = true }, true, true
xtex@xtex1 ~/s/z/zig (master)> zig version
0.15.0-dev.56+d0911786c
xtex@xtex1 ~/s/z/zig (master)> zig run repro.zig -fno-llvm
ret: repro.Data{ .rd = 1, .rl = false }, false, true
xtex@xtex1 ~/s/z/zig (master)> zig run repro.zig -fllvm
ret: repro.Data{ .rd = 1, .rl = true }, true, true

Behavior test: 1951 passed; 143 skipped; 0 failed.

[jacobly0]

Can we keep it simple with something like this, to reduce the impact on the soon-to-be-rewritten-anyway code:

--- a/src/arch/x86_64/CodeGen.zig
+++ b/src/arch/x86_64/CodeGen.zig
@@ -91810,12 +91810,15 @@ fn airStore(self: *CodeGen, inst: Air.Inst.Index, safety: bool) !void {
         const reg_locks = self.register_manager.lockRegsAssumeUnused(3, .{ .rdi, .rsi, .rcx });
         defer for (reg_locks) |lock| self.register_manager.unlockReg(lock);
 
+        const ptr_ty = self.typeOf(bin_op.lhs);
+        const ptr_info = ptr_ty.ptrInfo(zcu);
+        const is_packed = ptr_info.flags.vector_index != .none or ptr_info.packed_offset.host_size > 0;
+        if (is_packed) try self.spillEflagsIfOccupied();
+
         const src_mcv = try self.resolveInst(bin_op.rhs);
         const ptr_mcv = try self.resolveInst(bin_op.lhs);
-        const ptr_ty = self.typeOf(bin_op.lhs);
 
-        const ptr_info = ptr_ty.ptrInfo(zcu);
-        if (ptr_info.flags.vector_index != .none or ptr_info.packed_offset.host_size > 0) {
+        if (is_packed) {
             try self.packedStore(ptr_ty, ptr_mcv, src_mcv);
         } else {
             try self.store(ptr_ty, ptr_mcv, src_mcv, .{ .safety = safety });

This change also requires regression test coverage for the linked issues (in test/behavior, but not in x86_64) to be able to merge and close them.

@ifreund can you check the impact on #22790 when you get a chance to rule out any issues caused by this bug.

[maybe-xtex]

What happened in the aarch64 CI? The aarch64 debug build succeeded.

[ifreund]

@ifreund can you check the impact on #22790 when you get a chance to rule out any issues caused by this bug.

This indeed seems to fix the behavior described in #22790, I can't find any signs of miscompilation in river with this PR applied.

[jacobly0]

Can you add a test that doesn't have the & to cover #20113. Also just move them to packed-struct.zig while you are at it.

[xtexx]

Done and rebased.

[jacobly0]

Last nit I swear, since test names need to be unique, it should be something more like packed struct store of comparison result.